What happened?

Liquid AI has released LFM2.5-VL-3B, a 3.1-billion-parameter vision-language model built for on-device deployment. It reads digital screens across mobile, web and desktop, grounds objects to coordinates, parses documents and charts, and calls tools from text or image input.

The company reports an average of 69.4 across 28 vision benchmarks. That matches the 4.7-billion-parameter InternVL-3.5-4B and sits 0.7 points behind Qwen3.5-4B, also 4.7 billion. The model is non-reasoning: it answers directly, which keeps latency low.

What running on device means

The model fits in roughly 3 GB of memory and decodes 228 tokens per second on an Apple M5 Max. The checkpoint ships in four formats — native, GGUF, ONNX and MLX — and day-one runtimes include llama.cpp, MLX, vLLM, SGLang and ONNX.

The point here is not the size but where it runs. A model that reads your screen going to the cloud means everything on that screen travels over the network. A model running on the device removes that question entirely — and in regulated environments that is a compliance matter rather than a performance one.

What changed

  • Screen and UI understanding: an average of 80.7 on ScreenSpot-v2 — desktop 78.7, mobile 81.2, web 82.2. Liquid AI reports Gemma-4-E4B at 51.2 and Qwen3.5-4B at 78.5 on the same test, with the larger InternVL-3.5-4B ahead at 84.1.
  • Function calling: new to this line. ToolSandbox moves from 26.4 to 59.5 and BFCL v4 from 20.5 to 32.5.
  • Grounding: RefCOCO average precision@1 rises from 57.1 to 87.9, a 30-point gain the company attributes to scaled synthetic data.
  • Multi-image input: BLINK improves from 50.2 to 61.5 and MuirBench from 34.9 to 58.3.

Architecture and training

The language backbone is LFM2.5-2.6B and the vision tower is a SigLIP2 NaFlex shape-optimised 400M encoder. NaFlex handles native resolution by splitting large images into non-overlapping 512×512 patches plus a resized whole-image thumbnail. Context length is 32,768 tokens and 16 languages are supported.

Pre-training used approximately 34 trillion tokens. The vocabulary was doubled to 128K by extending the existing tokeniser in place, which improves coverage of non-Latin scripts. Post-training is supervised fine-tuning with knowledge distillation from a larger teacher, followed by multi-reward reinforcement learning.

The detail in the licence

The LFM Open License v1.0 is Apache-2.0-based with one change: free commercial use ends once a company's annual revenue reaches $10 million. Independent developers, startups and small and mid-sized businesses under that line can ship commercially at no cost; those above it must negotiate a commercial licence with Liquid AI. Research, education and non-profit use carry no revenue limit.

What is not settled

All the benchmark results rest on the company's own measurement, run with vLLM 0.26.0 in non-reasoning mode. There is no independent verification yet. That the weights are downloadable is what makes such verification possible — unlike a closed model's score claim, these numbers can be reproduced by someone else.